In this project we try to understand the issues that must be considered when implementing stream ciphers on superscalar processors with deep pipelines. One of the goals is to find ways to exploit the instruction level parallelism and to use all available bus and memory bandwidth to achieve the highest possible rate of encryption with copying.

The initial experiments demonstrate that blocks of data can be copied with encryption for virtually the same cost as copying alone. We have achieved these results by paying very careful attention to the scheduling of instructions and memory references, loop unrolling, and careful hand-optimization of the inner loop of both copy and copy-plus-encrypt operations. In other words, memory copy with encryption can be made virtually as efficient as a traditional memory copy.